Clean up voice recordings, podcasts and voice notes. Select a few seconds that contain only background noise, and the tool learns that noise and subtracts it from the whole file. Compare the original, the cleaned audio and the removed noise, then download a WAV. Nothing is uploaded.
No recording at hand?
Original Cleaned Noise sample
Noise sample (seconds)
to
Drag on the waveform across a part with only background noise: no voice, no music.
Noise sample Whole recording
0:00.0
Background noise—
Overall level—
Length—
Advertisement
Advertisement
How to use the Audio Noise Reducer
Open a recording: drop an MP3, WAV, M4A, OGG, FLAC or WEBM file on the box, or try the noisy sample.
Check the noise sample: the quietest part is selected automatically. If it contains speech or music, drag across a part with only background noise instead.
Press Reduce noise, then press Play and switch between Original, Cleaned and Removed to judge the result.
Fine-tune: raise Strength for more noise removal, lower Max reduction or raise Smoothing if voices sound watery or robotic.
Download WAV when you are happy with the sound.
Getting the cleanest result
Choose a clean noise sample. The sample should contain only the background sound you want removed: the pause before someone starts talking, or a gap between sentences. Even a single word inside it makes the tool treat speech as noise.
Start gentle. The defaults (60% strength, 18 dB maximum reduction) suit most voice recordings. Lowering the noise by 10 to 20 dB already sounds much cleaner; removing every trace of noise tends to make voices sound thin or watery.
Listen to Removed. It should sound like hiss or hum only. If you can follow the words in it, lower the Strength.
Use the high-pass filter for rumble. Traffic, desk bumps, wind and air conditioning put energy below 80 Hz, where voices have almost nothing. The filter removes it without touching speech.
Normalize afterwards. Noise reduction lowers the overall level slightly. Run the cleaned file through the Audio Normalizer to bring it to -16 LUFS for podcasts or -14 LUFS for video.
Settings explained
Setting
What it does
Turn it up when
Turn it down when
Strength
How much of the learned noise power is subtracted from each frequency band (oversubtraction factor 0 to 3).
Noise is still clearly audible
Speech sounds hollow, or words appear in Removed
Max reduction
The largest cut any band can get. 18 dB means noise can be lowered to about 1/8 of its level, never fully muted.
You need a quieter background
You hear chirpy "musical noise" or the background pumps
Smoothing
Averages the reduction across neighboring frequencies and lets it recover slowly over time.
The result sounds watery or robotic
Quiet word endings get swallowed
High-pass 80 Hz
A 12 dB per octave filter that removes deep rumble.
Recording has traffic, wind or handling noise
Music with deep bass, or a very deep voice
The method: spectral subtraction
For each short frame the tool computes the power spectrum |X(f)|² with a Fast Fourier Transform (Hann window, 75% overlap). With N(f) the average noise power learned from your sample and α the strength, each frequency band gets the gain G(f) = max(√(1 − α · N(f) / |X(f)|²), floor), where floor = 10−max reduction / 20. Bands far above the noise keep a gain close to 1, bands at the noise level are turned down to the floor. The gains are smoothed, applied, and the frames are transformed back and overlap-added into the cleaned audio. This classic method (Boll, 1979) is effective for steady noise and is the basis of many "noise reduction" effects in audio editors.
Frequently asked questions
How does the noise reduction work?
It uses spectral subtraction. The audio is split into short overlapping frames (about 43 ms at 48 kHz) and converted to frequencies with a Fast Fourier Transform. The average noise level in each frequency band is learned from the noise sample you select. In every frame, bands that are not much louder than the noise are turned down, while bands with speech or music well above the noise are kept. The frames are then turned back into sound and overlapped again.
What kind of noise can it remove?
Steady background noise works best: hiss from microphones and preamps, computer fans, air conditioning, electrical hum and buzz, traffic rumble and room tone. Sudden or changing sounds such as a dog barking, keyboard clicks, music or other people talking cannot be separated this way, because they do not match the learned noise profile.
Why does my cleaned audio sound watery or robotic?
These 'musical noise' artifacts appear when too much is removed. Lower the Strength, reduce Max reduction to 10 to 15 dB, or raise Smoothing. Leaving a little natural background noise usually sounds better than removing all of it. Also check that the noise sample contains no speech.
How long should the noise sample be?
Half a second to two seconds of pure background noise is ideal. The tool needs at least about 0.1 seconds. Recording a few seconds of silence at the start of every recording makes noise reduction much easier.
What does the Removed playback option do?
It plays the difference between the original and the cleaned audio, so you hear exactly what was taken out. If you can hear words clearly in it, the settings are too aggressive and parts of the voice are being removed.
Is my recording uploaded?
No. Decoding, analysis, noise reduction and WAV encoding all run in your browser with JavaScript and the Web Audio API. The file never leaves your device.
Can it remove noise from a video?
Extract the audio first with the Video to MP3 tool, clean it here, then put the cleaned audio back with the Add Audio to Video tool.